DOCSLIB.ORG

Critical Section and Mutual Exclusion

Total Page：16

File Type：pdf, Size：1020Kb

Download full-text PDF Read full-text
Critical Section and Mutual Exclusion Part III Synchronization Critical Section and Mutual Exclusion The question of whether computers can think is just like the question of whether submarines can swim 1 Fall 2015 Edsger W. Dijkstra Process Synchronization Topics . Why is synchronization needed? . Race Conditions . Critical Sections . Pure Software Solutions . Hardware Support . Semaphores . Race Conditions, Revisited . Monitors 2 Synchronization Needed! 1/6 int a[3] = { 3, 4, 5}; Process 1 Process 2 a[1] = a[0] + a[1]; a[2] = a[1] + a[2]; a[3] = { 3, ?, ? } Statement level execution interleaving 3 Synchronization Needed! 2/6 int a[3] = { 3, 4, 5}; Process 1 Process 2 a[1] = a[0] + a[1]; a[2] = a[1] + a[2]; .If process 1 updates a[1] first, a[1] is 7, and a[ ]={3,7,5} .Then, process 2 uses the new a[1] to computes a[2], and a[ ]={3,7,12} .If process 2 uses a[1] first, now a[2] is 9, and a[ ]={3,4,9} .Then, process 1 computes a[1], and a[ ]={3,7,9} Results are non-deterministic! 4 Synchronization Needed! 3/6 int Count = 10; Higher-level language statements are not atomic Process 1 Process 2 Count++; Count--; Count = 9, 10 or 11? 5 Synchronization Needed! 4/6 int Count = 10; Process 1 Process 2 LOAD Reg, Count LOAD Reg, Count ADD #1 SUB #1 STORE Reg, Count STORE Reg, Count The problem is that the execution flow may be switched in the middle. Results become non-deterministic! instruction level execution interleaving 6 Synchronization Needed! 5/6 Process 1 Process 2 Inst Reg Memory Inst Reg Memory LOAD 10 10 LOAD 10 10 SUB 9 10 ADD 11 10 STORE 11 11 overwrites the previous value 11 STORE 9 9 Always use instruction level interleaving to show race conditions 7 Synchronization Needed! 6/6 Process 1 Process 2 Inst Reg Memory Inst Reg Memory LOAD 10 10 ADD 11 10 LOAD 10 10 SUB 9 10 STORE 9 9 STORE 11 11 overwrites the previous value 9 Always use instruction level interleaving to show race conditions 8 Race Conditions . A Race Condition occurs, if two or more processes/threads manipulate a shared resource concurrently, and the outcome of the execution depends on the particular order in which the access takes place. Synchronization is needed to prevent race conditions from happening. Synchronization is a difficult topic. Don’t miss classes; otherwise, you will miss a lot of things. 9 Execution Sequence Notes . You should use instruction level interleaving to demonstrate the existence of race conditions, because a) higher-level language statements are not atomic and can be switched in the middle of execution b) instruction level interleaving can show clearly the “sharing” of a resource among processes and threads. 10 Critical Section . A critical section, CS, is a section of code in which a process accesses shared resources. int count; // shared process 1 process 2 process 3 count++; count--; cout << count; These are critical sections since count is a shared resource 11 Mutual Exclusion . To avoid race conditions, the execution of critical sections must be mutually exclusive (e.g., at most one process can be in its critical section at any time). The critical-section problem is to design a protocol with which processes can use to cooperate and ensure mutual exclusion. 12 The Critical Section Protocol do { . A critical section protocol consists of two parts: an entry section and an exit entry section section. critical section . Between them is the exit section critical section that must run in a mutually exclusive way. } while (1); 13 Good Solutions to the CS Problem . A good solution to the critical section problem must satisfy the following three conditions: Mutual Exclusion Progress Bounded Waiting . Moreover, the solution cannot depend on CPU’s relative speed, timing, scheduling policy and other external factors. 14 Mutual Exclusion . If a process P is executing in its critical section, no other processes can be executing in their corresponding critical sections. The entry protocol should be able to block processes that wish to enter but cannot. When the process that is executing in its critical section exits, the entry protocol must be able to know this fact and allows a waiting process to enter. 15 Progress . If no process is executing in its critical section and some processes want to enter their corresponding critical sections, then 1. Only those processes that are waiting to enter can participate in the competition (to enter their critical sections) and no other processes can influence this decision. 2. This decision cannot be postponed indefinitely (i.e., finite decision time). Thus, one of the waiting processes can enter its critical section. 16 Bounded Waiting . After a process made a request to enter its critical section and before it is granted the permission to enter, there exists a bound on the number of turns that other processes are allowed to enter. Finite is not the same as bounded. The former means any value you can write down (e.g., billion, trillion, etc) while the latter means this value has to be no larger than a particular one (i.e., the bound). 17 Progress vs. Bounded Waiting . Progress does not imply Bounded Waiting: Progress says a process can enter with a finite decision time. It does not say which process can enter, and there is no guarantee for bounded waiting. Bounded Waiting does not imply Progress: Even through we have a bound, all processes may be locked up in the enter section (i.e., failure of Progress). Therefore, Progress and Bounded Waiting are independent of each other. 18 A Few Related Terms: 1/7 . Deadlock-Freedom: If two or more processes are trying to enter their critical sections, one of them will eventually enter. This is Progress without the “outsiders having no influence” condition. Since the enter section is able to select a process to enter, the decision time is certainly finite. 19 A Few Related Terms: 2/7 . r-Bounded Waiting: There exists a fixed value r such that after a process made a request to enter its critical section and before it is granted the permission to enter, no more than r other processes are allowed to enter. Therefore, bounded waiting means there is a r such that the waiting is r-bounded. 20 A Few Related Terms: 3/7 . FIFO: No process that is about to enter its critical section can pass an already waiting process. FIFO is usually referred to as 0- bounded. Linear-Waiting (1-Bounded Waiting): No process can enter its critical section twice while there is a process waiting. 21 A Few Related Terms: 4/7 . Starvation-Freedom: If a process is trying to enter its critical section, it will eventually enter. Questions: 1. Does starvation-freedom imply deadlock-freedom? 2. Does starvation-freedom imply bounded-waiting? 3. Does bounded-waiting imply starvation-freedom? 4. Does bounded-waiting AND deadlock-freedom imply starvation-freedom? 22 A Few Related Terms: 5/7 . Question (1): Does starvation-freedom imply deadlock-freedom? . Yes! If every process can eventually enter its critical section, although waiting time may vary, it means the decision time of selecting a process is finite. Otherwise, all processes would wait in the enter section. 23 A Few Related Terms: 6/7 . Question (2): Does starvation-freedom imply bounded-waiting? . No! This is because the waiting time may not be bounded even though each process can enter its critical section. 24 A Few Related Terms: 7/7 . Question (3): Does bounded-waiting imply starvation-freedom? . No. Bounded-Waiting does not say if a process can actually enter. It only says there is a bound. For example, all processes are locked up in the enter section (i.e., failure of Progress). We need Progress + Bounded-Waiting to imply Starvation-Freedom (Question (4)). In fact, Progress + Bounded-Waiting is stronger than Starvation-Freedom. Why? 25 The End 26 .
Load more

Recommended publications
  • Introduction to Multi-Threading and Vectorization Matti Kortelainen Larsoft Workshop 2019 25 June 2019 Outline
    Introduction to multi-threading and vectorization Matti Kortelainen LArSoft Workshop 2019 25 June 2019 Outline Broad introductory overview: • Why multithread? • What is a thread? • Some threading models – std::thread – OpenMP (fork-join) – Intel Threading Building Blocks (TBB) (tasks) • Race condition, critical region, mutual exclusion, deadlock • Vectorization (SIMD) 2 6/25/19 Matti Kortelainen | Introduction to multi-threading and vectorization Motivations for multithreading Image courtesy of K. Rupp 3 6/25/19 Matti Kortelainen | Introduction to multi-threading and vectorization Motivations for multithreading • One process on a node: speedups from parallelizing parts of the programs – Any problem can get speedup if the threads can cooperate on • same core (sharing L1 cache) • L2 cache (may be shared among small number of cores) • Fully loaded node: save memory and other resources – Threads can share objects -> N threads can use significantly less memory than N processes • If smallest chunk of data is so big that only one fits in memory at a time, is there any other option? 4 6/25/19 Matti Kortelainen | Introduction to multi-threading and vectorization What is a (software) thread? (in POSIX/Linux) • “Smallest sequence of programmed instructions that can be managed independently by a scheduler” [Wikipedia] • A thread has its own – Program counter – Registers – Stack – Thread-local memory (better to avoid in general) • Threads of a process share everything else, e.g. – Program code, constants – Heap memory – Network connections – File handles
    [Show full text]
  • CSC 553 Operating Systems Multiple Processes
    CSC 553 Operating Systems Lecture 4 - Concurrency: Mutual Exclusion and Synchronization Multiple Processes • Operating System design is concerned with the management of processes and threads: • Multiprogramming • Multiprocessing • Distributed Processing Concurrency Arises in Three Different Contexts: • Multiple Applications – invented to allow processing time to be shared among active applications • Structured Applications – extension of modular design and structured programming • Operating System Structure – OS themselves implemented as a set of processes or threads Key Terms Related to Concurrency Principles of Concurrency • Interleaving and overlapping • can be viewed as examples of concurrent processing • both present the same problems • Uniprocessor – the relative speed of execution of processes cannot be predicted • depends on activities of other processes • the way the OS handles interrupts • scheduling policies of the OS Difficulties of Concurrency • Sharing of global resources • Difficult for the OS to manage the allocation of resources optimally • Difficult to locate programming errors as results are not deterministic and reproducible Race Condition • Occurs when multiple processes or threads read and write data items • The final result depends on the order of execution – the “loser” of the race is the process that updates last and will determine the final value of the variable Operating System Concerns • Design and management issues raised by the existence of concurrency: • The OS must: – be able to keep track of various processes
    [Show full text]
  • Towards Gradually Typed Capabilities in the Pi-Calculus
    Towards Gradually Typed Capabilities in the Pi-Calculus Matteo Cimini University of Massachusetts Lowell Lowell, MA, USA matteo [email protected] Gradual typing is an approach to integrating static and dynamic typing within the same language, and puts the programmer in control of which regions of code are type checked at compile-time and which are type checked at run-time. In this paper, we focus on the π-calculus equipped with types for the modeling of input-output capabilities of channels. We present our preliminary work towards a gradually typed version of this calculus. We present a type system, a cast insertion procedure that automatically inserts run-time checks, and an operational semantics of a π-calculus that handles casts on channels. Although we do not claim any theoretical results on our formulations, we demonstrate our calculus with an example and discuss our future plans. 1 Introduction This paper presents preliminary work on integrating dynamically typed features into a typed π-calculus. Pierce and Sangiorgi have defined a typed π-calculus in which channels can be assigned types that express input and output capabilities [16]. Channels can be declared to be input-only, output-only, or that can be used for both input and output operations. As a consequence, this type system can detect unintended or malicious channel misuses at compile-time. The static typing nature of the type system prevents the modeling of scenarios in which the input- output discipline of a channel is discovered at run-time. In these scenarios, such a discipline must be enforced with run-time type checks in the style of dynamic typing.
    [Show full text]
  • Enhanced Debugging for Data Races in Parallel
    ENHANCED DEBUGGING FOR DATA RACES IN PARALLEL PROGRAMS USING OPENMP A Thesis Presented to the Faculty of the Department of Computer Science University of Houston In Partial Fulfillment of the Requirements for the Degree Master of Science By Marcus W. Hervey December 2016 ENHANCED DEBUGGING FOR DATA RACES IN PARALLEL PROGRAMS USING OPENMP Marcus W. Hervey APPROVED: Dr. Edgar Gabriel, Chairman Dept. of Computer Science Dr. Shishir Shah Dept. of Computer Science Dr. Barbara Chapman Dept. of Computer Science, Stony Brook University Dean, College of Natural Sciences and Mathematics ii iii Acknowledgements A special thank you to Dr. Barbara Chapman, Ph.D., Dr. Edgar Gabriel, Ph.D., Dr. Shishir Shah, Ph.D., Dr. Lei Huang, Ph.D., Dr. Chunhua Liao, Ph.D., Dr. Laksano Adhianto, Ph.D., and Dr. Oscar Hernandez, Ph.D. for their guidance and support throughout this endeavor. I would also like to thank Van Bui, Deepak Eachempati, James LaGrone, and Cody Addison for their friendship and teamwork, without which this endeavor would have never been accomplished. I dedicate this thesis to my parents (Billy and Olivia Hervey) who have always chal- lenged me to be my best, and to my wife and kids for their love and sacrifice throughout this process. ”Our greatest weakness lies in giving up. The most certain way to succeed is always to try just one more time.” – Thomas Alva Edison iv ENHANCED DEBUGGING FOR DATA RACES IN PARALLEL PROGRAMS USING OPENMP An Abstract of a Thesis Presented to the Faculty of the Department of Computer Science University of Houston In Partial Fulfillment of the Requirements for the Degree Master of Science By Marcus W.
    [Show full text]
  • Reconciling Real and Stochastic Time: the Need for Probabilistic Refinement
    DOI 10.1007/s00165-012-0230-y The Author(s) © 2012. This article is published with open access at Springerlink.com Formal Aspects Formal Aspects of Computing (2012) 24: 497–518 of Computing Reconciling real and stochastic time: the need for probabilistic refinement J. Markovski1,P.R.D’Argenio2,J.C.M.Baeten1,3 and E. P. de Vink1,3 1 Eindhoven University of Technology, Eindhoven, The Netherlands. E-mail: [email protected] 2 FaMAF, Universidad Nacional de Cordoba,´ Cordoba,´ Argentina 3 Centrum Wiskunde & Informatica, Amsterdam, The Netherlands Abstract. We conservatively extend an ACP-style discrete-time process theory with discrete stochastic delays. The semantics of the timed delays relies on time additivity and time determinism, which are properties that enable us to merge subsequent timed delays and to impose their synchronous expiration. Stochastic delays, however, interact with respect to a so-called race condition that determines the set of delays that expire first, which is guided by an (implicit) probabilistic choice. The race condition precludes the property of time additivity as the merger of stochastic delays alters this probabilistic behavior. To this end, we resolve the race condition using conditionally-distributed unit delays. We give a sound and ground-complete axiomatization of the process the- ory comprising the standard set of ACP-style operators. In this generalized setting, the alternative composition is no longer associative, so we have to resort to special normal forms that explicitly resolve the underlying race condition. Our treatment succeeds in the initial challenge to conservatively extend standard time with stochastic time. However, the ‘dissection’ of the stochastic delays to conditionally-distributed unit delays comes at a price, as we can no longer relate the resolved race condition to the original stochastic delays.
    [Show full text]
  • Lecture 18 18.1 Distributed and Concurrent Systems 18.2 Race
    CMPSCI 691ST Systems Fall 2011 Lecture 18 Lecturer: Emery Berger Scribe: Sean Barker 18.1 Distributed and Concurrent Systems In this lecture we discussed two broad issues in concurrent systems: clocks and race detection. The first paper we read introduced the Lamport clock, which is a scalar providing a logical clock value for each process. Every time a process sends a message, it increments its clock value, and every time a process receives a message, it sets its clock value to the max of the current value and the message value. Clock values are used to provide a `happens-before' relationship, which provides a consistent partial ordering of events across processes. Lamport clocks were later extended to vector clocks, in which each of n processes maintains an array (or vector) of n clock values (one for each process) rather than just a single value as in Lamport clocks. Vector clocks are much more precise than Lamport clocks, since they can check whether intermediate events from anyone have occurred (by looking at the max across the entire vector), and have essentially replaced Lamport clocks today. Moreover, since distributed systems are already sending many messages, the cost of maintaining vector clocks is generally not a big deal. Clock synchronization is still a hard problem, however, because physical clock values will always have some amount of drift. The FastTrack paper we read leverages vector clocks to do race detection. 18.2 Race Detection Broadly speaking, a race condition occurs when at least two threads are accessing a variable, and one of them writes the value while another is reading.
    [Show full text]
  • Breadth First Search Vectorization on the Intel Xeon Phi
    Breadth First Search Vectorization on the Intel Xeon Phi Mireya Paredes Graham Riley Mikel Luján University of Manchester University of Manchester University of Manchester Oxford Road, M13 9PL Oxford Road, M13 9PL Oxford Road, M13 9PL Manchester, UK Manchester, UK Manchester, UK [email protected] [email protected] [email protected] ABSTRACT Today, scientific experiments in many domains, and large Breadth First Search (BFS) is a building block for graph organizations can generate huge amounts of data, e.g. social algorithms and has recently been used for large scale anal- networks, health-care records and bio-informatics applica- ysis of information in a variety of applications including so- tions [8]. Graphs seem to be a good match for important cial networks, graph databases and web searching. Due to and large dataset analysis as they can abstract real world its importance, a number of different parallel programming networks. To process large graphs, different techniques have models and architectures have been exploited to optimize been applied, including parallel programming. The main the BFS. However, due to the irregular memory access pat- challenge in the parallelization of graph processing is that terns and the unstructured nature of the large graphs, its large graphs are often unstructured and highly irregular and efficient parallelization is a challenge. The Xeon Phi is a this limits their scalability and performance when executing massively parallel architecture available as an off-the-shelf on off-the-shelf systems [18]. accelerator, which includes a powerful 512 bit vector unit The Breadth First Search (BFS) is a building block of with optimized scatter and gather functions.
    [Show full text]
  • Rapid Prototyping of High-Performance Concurrent Java Applications
    Rapid Prototyping of High-Performance Concurrent Java Applications K.J.R. Powell Master of Science School of Computer Science School of Informatics University of Edinburgh 2002 (Graduation date:December 2002) Abstract The design and implementation of concurrent applications is more challenging than that of sequential applications. The aim of this project is to address this challenge by producing an application which can generate skeleton Java systems from a high-level PEPA modelling language description. By automating the process of translating the design into a Java skeleton system, the result will maintain the performance and be- havioural characteristics of the model, providing a sound framework for completing the concurrent application. This method accelerates the process of initial implementation whilst ensuring the system characterisics remain true to the high-level model. i Acknowledgements Many thanks to my supervisor Stephen Gilmore for his excellent guidance and ideas. I’d also like to express my gratitude to L.L. for his assistance above and beyond the call of duty in the field of editing. ii Declaration I declare that this thesis was composed by myself, that the work contained herein is my own except where explicitly stated otherwise in the text, and that this work has not been submitted for any other degree or professional qualification except as specified. (K.J.R. Powell) iii Table of Contents 1 Introduction 1 1.1 Principal Goal . 1 1.2 PEPA and Java . 2 1.3 Project Objective . 2 2 Performance Evaluation Process Algebra 4 2.1 Overview of PEPA and its role in design . 4 2.2 The Syntax of PEPA .
    [Show full text]
  • Towards Systematic Black-Box Testing for Exploitable Race Conditions in Web Apps
    1 Faculty of Electrical Engineering, Mathematics & Computer Science Towards Systematic Black-Box Testing for Exploitable Race Conditions in Web Apps Rob J. van Emous [email protected] Master Thesis June 2019 Supervisors: prof. dr. M. Huisman dr. ing. E. Tews (Computest) M.Sc. D. Keuper Formal Methods and Tools Group Faculty of Electrical Engineering, Mathematics and Computer Science University of Twente P.O. Box 217 7500 AE Enschede The Netherlands ABSTRACT As web applications get more complicated and more integrated into our daily lives, securing them against known cyber attacks is of great importance. Performing se- curity tests is the primary way to discover the current risks and act accordingly. In order to perform these test in a systematic way, testers create and use numer- ous vulnerability lists and testing guidelines. Often, dedicated security institutions like Escal Institute of Advanced Technologies (SANS), MITRE, Certified Secure, and the Open Web Application Security Project (OWASP) create these guidelines. These lists are not meant to be exhaustive, but as the introduction in the Common Weakness Enumeration (CWE) of MITRE and SANS by Martin et al. (2011) puts it: they "(..) evaluated each weakness based on prevalence, importance, and the likelihood of exploit". This seems like a reasonable way of thinking, but as we have shown in this research, it can also lead to oversight of essential, but stealthy security issues. In this research, we will focus on one such stealthy issue in web apps called a race condition. Race conditions are known for a very long time as research by Abbott et al.
    [Show full text]
  • Iota: a Calculus for Internet of Things Automation
    Iota: A Calculus for Internet of Things Automation Julie L. Newcomb Satish Chandra∗ Jean-Baptiste Jeannin† Paul G. Allen School of Computer Samsung Research America Samsung Research America Science & Engineering Mountain View, CA, USA Mountain View, CA, USA University of Washington schandra@acm:org jb:jeannin@gmail:com Seattle, WA, USA newcombj@cs:washington:edu Cole Schlesinger‡ Manu Sridharan§ Samsung Research America Samsung Research America Mountain View, CA, USA Mountain View, CA, USA cole@schlesinger:tech manu@sridharan:net Abstract analyses from the software engineering literature, and ex- Programmatically controllable home devices are proliferat- pressive by encoding sixteen programs from these home ing, ranging from lights, locks, and motion sensors to smart automation platforms. Along the way, we highlight how the refrigerators, televisions, and cameras, giving end users un- design of the Iota semantics rules out subtle classes of bugs. precedented control over their environment. New domain- CCS Concepts • Software and its engineering → Gen- specific languages are emerging to supplant general purpose eral programming languages; • Social and professional programming platforms as a means for end users to con- topics → History of programming languages; figure home automation. These languages, based on event- condition-action (ECA) rules, have an appealing simplicity. Keywords domain-specific languages, language design and But programmatic control allows users write to programs implementation with bugs, introducing the frustrations of software engineer- ACM Reference Format: ing with none of the tool support. The subtle semantics of Julie L. Newcomb, Satish Chandra, Jean-Baptiste Jeannin, Cole the home automation domain—as well as the varying in- Schlesinger, and Manu Sridharan.
    [Show full text]
  • Preview – Interprocess Communication
    Preview – Interprocess Communication Race Condition Mutual Exclusion with Busy Waiting Disabling Interrupts Lock Variables Strict Alternation Peterson’s Solution Test and Set Lock – help from hardware Sleep and Wakeup The Producer-Consumer Problem Race Condition in Producer-Consumer Problem Semaphores The producer-consumer problem using semaphores Mutexes Monitors CS 431 Operating Systems 1 Interprocess Communication Critical section (critical region) – The part of program where the shared memory is accessed. Three issues in InterProcess Communication (IPC) 1. How one process can pass information to another 2. How to make sure two or more processes do not get into the critical region at the same time 3. Proper sequencing when dependencies are present (e.g. A creates outputs, B consumes the outputs) CS 431 Operating Systems 2 Race Condition A situation where two or more processes are reading or writing some shared data and the final result depends on who runs precisely when, is called race condition. Network Printer Spooler Directory CS 431 Operating Systems 3 Race Condition Process B will never Process A tries to send a job to receive any output from spooler, Process A reads the printer !! in = 7, process A times out and goes to “ready” state before updating in = in + 1. CPU switches to Process B. Process B also tries to send a job to spooler. Process B reads in = 7, writes its job name in slot 7, updates i to 8 and then goes to “ready” state. CPU switches to Process A. Process A starts from the place it left off. Process writes its job name in slot 7, update i to 8.
    [Show full text]
  • Comparison of Modeling Approaches to Business Process Performance Evaluation Kelly Rosa Braghetto, João Eduardo Ferreira, Jean-Marc Vincent
    Comparison of Modeling Approaches to Business Process Performance Evaluation Kelly Rosa Braghetto, João Eduardo Ferreira, Jean-Marc Vincent To cite this version: Kelly Rosa Braghetto, João Eduardo Ferreira, Jean-Marc Vincent. Comparison of Modeling Ap- proaches to Business Process Performance Evaluation. [Research Report] RR-7065, INRIA. 2009. inria-00424452 HAL Id: inria-00424452 https://hal.inria.fr/inria-00424452 Submitted on 15 Oct 2009 HAL is a multi-disciplinary open access L’archive ouverte pluridisciplinaire HAL, est archive for the deposit and dissemination of sci- destinée au dépôt et à la diffusion de documents entific research documents, whether they are pub- scientifiques de niveau recherche, publiés ou non, lished or not. The documents may come from émanant des établissements d’enseignement et de teaching and research institutions in France or recherche français ou étrangers, des laboratoires abroad, or from public or private research centers. publics ou privés. INSTITUT NATIONAL DE RECHERCHE EN INFORMATIQUE ET EN AUTOMATIQUE Comparison of Modeling Approaches to Business Process Performance Evaluation Kelly Rosa Braghetto — João Eduardo Ferreira — Jean-Marc Vincent N° 7065 October 2009 apport de recherche ISSN 0249-6399 ISRN INRIA/RR--7065--FR+ENG Comparison of Modeling Approaches to Business Process Performance Evaluation Kelly Rosa Braghetto ∗ † , João Eduardo Ferreira ∗ , Jean-Marc Vincent ‡ Thème : Calcule distribué et applications à très haute performance Équipe-Projet MESCAL Rapport de recherche n° 7065 — October
    [Show full text]